RADIOENGINEERING, VOL. 22, NO. 1, APRIL 2013
371
A Large-scale RF-based Indoor Localization System
Using Low-complexity Gaussian Filter and Improved
Bayesian Inference
Long XIAO1, Ye YIN1, Xiangnong WU1, Jianwei WANG2
1
College of Information, Mechanical and Electrical Engineering, Shanghai Normal University, Shanghai, 200234, China
2
Shanghai Zhen Zhuo Electrical Technology Ltd. Co., Shanghai, 201100, China
loongshaw@gmail.com, yinye@188.com, xnwu@shnu.edu.cn, wangjianwei@zenkore.com
Abstract. The growing convergence among mobile computing device and smart sensors boosts the development of
ubiquitous computing and smart spaces, where localization
is an essential part to realize the big vision. The general
localization methods based on GPS and cellular techniques
are not suitable for tracking numerous small size and limited power objects in the indoor case. In this paper, we
propose and demonstrate a new localization method, this
method is an easy-setup and cost-effective indoor localization system based on off-the-shelf active RFID technology.
Our system is not only compatible with the future smart
spaces and ubiquitous computing systems, but also suitable
for large-scale indoor localization. The use of low-complexity Gaussian Filter (GF), Wheel Graph Model (WGM)
and Probabilistic Localization Algorithm (PLA) make the
proposed algorithm robust and suitable for large-scale indoor positioning from uncertainty, self-adjective to varying
indoor environment. Using MATLAB simulation, we study
the system performances, especially the dependence on
a number of system and environment parameters, and their
statistical properties. The simulation results prove that our
proposed system is an accurate and cost-effective candidate for indoor localization.
Keywords
Indoor localization, Gaussian filter, wheel graph,
probabilistic localization algorithm.
1. Introduction
The growing of wireless communications, portable
computers, and smart sensors has generated much commercial and research interests in statistical methods to track
people and things inside stores, hospitals, warehouses, and
factories, where Global Positioning System (GPS) [1] devices generally do not work and Indoor Positioning System
(IPS) [2], [3], [4] aims to provide location estimation for
wireless devices, such as handheld devices and electronic
badges. In this paper, we propose and study a new method
for indoor localization, which is cost-effective for tracking
objects.
Nowadays, due to advantages such as small size, low
power, low cost and easy deployment, the Radio Frequency
Identification (RFID) [5] sensors are widely used to implement ubiquitous computing and smart space. We have successfully applied this technique to the 2010 Shanghai
World Expo [6]. With the capability of providing RSS
(Received Signal Strength) information, current advanced
RFID systems are one of the potential candidates for indoor
localization. Several RFID based systems have been proposed for tracking objects in indoor environments. As we
know, SpotON [2], [7] uses an aggregation algorithm for
three-dimensional localization. The tags use RSS information to obtain inter-tag distances based on empirical mapping between the two. SpotON assumes deterministic mapping between RSS and distance and does not account for
the range measurement uncertainty caused by the varying
environment. LANDMARC [2], [8] utilizes RSS measurement information to locate objects using k nearest reference
tags. To diminish the uncertainty of the detected range
caused by the varying environments, a large number of
reference tags must be distributed in the environment. This
limits its applications for most indoor scenarios, such as
people tracking, fire disaster site, and so on.
In this paper, we present a new method based on
off-the-shelf active RFID technology, Gaussian Filter (GF),
Wheel Graph Model (WGM) and Probabilistic Localization
Algorithm (PLA). It is easy to deploy and cost-effective for
indoor localization. Considering the uncertainties caused
by the varying environment, we incorporate a probabilistic
scheme based on Gaussian filter pretreatment, on-site calibration and Bayesian inference to improve the localization
accuracy. Bayesian inference was also used for traditional
cellular and WLAN-based localization system. But with an
additional WGM, our proposed system can be applied to
large-scale indoor localization.
The rest of the paper is organized as follows: Section
2 describes our method, including the basic system architecture, system overview, GF pretreatment, path loss model,
WGM and the PLA. In Section 3, the method of performance metric is introduced, followed by simulation results
and discussions. And finally in Section 4 we conclude the
paper.
372
LONG XIAO, YE YIN, XIANGNONG WU, JIANWEI WANG, A LARGE-SCALE RF-BASED INDOOR LOCALIZATION SYSTEM …
2. System Architecture
In our localization system the off-the-shelf long distance active RFID system [9] is used. The system works at
the range of 2.4 GHz frequency [10], with a minimal range
of 0.5 meter and maximum range of 80 meters. The reader
can receive Received Signal Strength Indicator (RSSI)
from every tag within its range. Each reader can detect up
to 200 tags simultaneously. Each RFID tag is pre-programmed with a unique 9-character ID (Identity) for identification by readers. In the next section, we will discuss the network layout of the system.
2.1 Network Layout
The system consists of three network layers: sensing
network layer, data collection layer, processing layer. The
sensing network layer is used to measure the RSSI information from the readers to objects (active tags) and to
transmit the information to next layer. The data collection
layer is used to collect RSSI information and to transmit
the data to next layer. The processing layer receives RSSI
information and processes the location information. In
practice, the whole detection area may be covered by several servers. For simplicity, in Fig. 1, we show the hierarchical architecture within the coverage of one server. Fig. 4
shows the reader and tags used in our experiments.
ers. The whole detection area is covered by the data collection layer. Every object tag should be within the readable
range of readers. The principle of readers’ deployment
should be satisfied that the distribution of the readers is not
in one line in the space, and all readers’ locations are
known. A reader is responsible for: 1) Collecting and decoding the signals emitted by the active tag in its coverage;
2) Measuring the RSSI for each tag within its range; 3) Reporting tag ID, corresponding RSSI, and its own ID number to the server. To realize these functions, each reader is
equipped with two interfaces: a RF interface that detects
tags within its range, and a communication interface that
transmits data to servers.
Servers: Each reader should be within the reach of at
least one server. We can see this from Fig. 1. Readers communicate the measured RSSIs of the tags with the server.
A server is responsible for: 1) Collection of RSSIs and IDs
coming from readers. 2) Calculating the location of the
object tags according to the positioning algorithm.
Reader
Tag
working frequency
2.4 GHz
modulation mode
MSK
communication distance
0.5 meter~80 meter
communication rate
250 kbps
working voltage
12 V
1.5 V
communication interface
RJ45, RS232, RS485
-
working temperature
-40°C~+80°C
Manufacturer: Shanghai Zhen Zhuo Electrical Technology Ltd. Co
Tab. 1. Manufacturer and parameters of reader and tag.
Fig. 2-4 show readers and tags used in the experiment,
Tab. 1 gives the Manufacturer and parameters of them.
Fig. 1. System structure.
Fig. 2. Reader.
Fig. 3. Tags.
Fig. 4. Reader and tags.
RFID tags: We use the active tags in our system, as
shown in Fig. 3. They are deployed in the sensing network
layer, and are divided into two categories: object tags and
reference tags. Each tracked object will be attached with
a unique active RFID tag, called "object tag", used for
identifying and tracking objects. Each tag has a unique ID,
hence, we can distinguish objects by the corresponding ID
number. Reference tags are the active RFID tags used for
calibrating environment parameters. The active tags will
periodically emit signals with their IDs.
RFID readers: The data collection layer consists of
readers. Each reader, as shown in Fig. 2, also has a unique
ID number. Every small detection area contains three read-
The principles of the system are as follows. After the
system is setup, the RFID readers will begin to detect signals sent by RFID tags. If any tag is located within its coverage, the reader will collect and report the RSSIs of the tag
and IDs both of the tag and the reader to the server. All the
data reported by the RFID readers will be processed at the
server. The server will estimate the locations of the objects
according to the localization algorithm using the RSS information and the readers’ locations.
2.2 Gaussian Filter Pretreatment
As we know, RSSI is very sensitive to the environment. This will limit the accuracy of the measurement. In
fact, the relationship between the signal strength and the
distance is not very clear, and there is a lot of volatility of
RSSI in the environment. During a certain period of time,
the RFID readers can read a series of RSSI values. However, the non-line-of-sight and multipath effects cause these
RSSI values with a lot of volatility. Hence, filtering of
RSSI is necessary to obtain a group of highly credible RSSI
values.
RADIOENGINEERING, VOL. 22, NO. 1, APRIL 2013
373
In the data collection stage, due to the volatility of
RSSI, we usually repeat the experiments. Supposing that
each experiment is independent, we can assume that the
distribution of RSSI values is a normal distribution.
Through Gaussian filter [11] processing, we can obtain
a set of relatively smooth values of RSSI with improved
reliability, and improved positioning accuracy.
Due to the RSSIs’ distribution is a normal distribution
model, its mean is μ, and standard deviation is σ. For a specific set of RSSI values, expressed as {RSSIt,t = 1,2,3,…n},
which is received at a fixed position from time slot 1 to slot
n. Assume that x  RSSIt, we use f(x) to represent Probability Density Function (PDF) of x, given by:
   x   2 
1
f x  
exp 

2
 2

 2
(1)
where n is the times of repeat measurements,

1 n
 RSSI t ,  
n t 1

1 n
 RSSI t  
n  1 t 1

2
.
We choose the high probability zone as P > 0.6 [12], which
is an experiment value. Through the above analysis, the
select interval of RSSI is [0.26σ + μ, 3.09σ + μ]. Fig. 5
gives the PDF and probability distribution of one set of
RSSIs’ sampling value.
Fig. 6 and Fig. 7 denote two sets of RSSI values for
which the distances from tags to readers are 2 meters and
3 meters, respectively. The number n of RSSI values in
each set is 200.
Above two kinds of circumstances each have two
hundred set of values, these values are unfiltered data. Our
filtering process is to eliminate those values which do not
meet the interval [0.26σ + μ, 3.09σ + μ]. This interval is to
use unfiltered data calculated out. After filtering, the rest of
data is filtered data.
We calculate respectively the arithmetic mean, its
estimated distance and estimation error of the two sets of
values. Tab. 2 shows the result of the calculation. Tab. 2
shows the comparison of the estimated distance in two
cases: filtered data and unfiltered data. From Tab. 2, we can
see that the reliability of RSSI has improved significantly
after Gaussian filtering.
Actual distance
(meter)
Mean of RSSI Estimated distance Estimation
(dBm)
(meter)
error
Unfiltered data
2.0
-67.0073
1.744379
12.78%
3.0
-72.4887
2.656766
11.44%
Filtered data
2.0
-68.2239
1.915108
4.24%
3.0
-73.7295
2.922223
2.59%
Tab. 2. Comparisons of two types of data.
2.3 Probability Model of Path Loss
We choose log-distance path loss model [13] as ideal
free space propagation path loss formula:
Fig. 5. RSSIs’ probability distribution function (PDF).
PLd   PLd 0   10n logd / d 0  .
(2)
In (2), PL(d) is propagation path loss, d is the actual distance from tags to readers, d0 is reference distance and usually the value of d0 is 1, n is the path loss exponent of the
environment.
However, there are a lot of interferences from outside
in the practical propagation, for example, walls, elevators,
furniture, human activities, multipath effects, shadow effects, etc. Hence equation (2) can be modified as:
Fig. 6. RSSI Sampled values (2 meter).
PL(d )  PL(d 0 )  10n log(d d 0 )  N (0,  2 )
(3)
where N(0, σ) denotes a Gaussian distribution with mean of
0 and variance of σ. Assume that Pr is the received power,
Pt is the constant transmit power, we have
Pr  Pt  PL(d ) .
(4)
From the above, d0 = 1, the expression 10n log(d/d0)
in (3) changes to 10n log(d), so (3) changes to (5)
Fig. 7. RSSI Sampled values (3 meter).
Pr  Pt  PL(d0 )  10n log(d )  N (0, 2 ) .
(5)
374
LONG XIAO, YE YIN, XIANGNONG WU, JIANWEI WANG, A LARGE-SCALE RF-BASED INDOOR LOCALIZATION SYSTEM …
As we know, Pt - PL(d0) is a constant, so we can set it
as a constant of α. PL(d0) = 10n log(d0), so 10n log(d0) is
a part of constant α.
Then, equation (5) can be changed as:
Pr    10n log d  N (0,  2 )
(6)
where Pr is RSSI, it could be regarded as a Gaussian distribution with the mean of (α - 10nlogd) and variance of σ2.
The PDF of RSSI is given by:
p Pr  RSSI 
  ( RSSI    10n log d ) 2  (7)

exp
.
2 2
 2


1
Fig. 9. Each small area is covered by three readers and each
reader’s location is known. Tags will periodically broadcast
the information to all readers around them, readers will
collect information (all tags ID and RSSI) from tags and
upload data (all readers ID, tags ID and RSSI) to the server.
Then, the server processes the uploaded data. The data include three types of number, if the data can’t be effectively
paired, it is very easy to create chaos in the post-processing
stage, i.e., we can’t distinguish which RSSI is uploaded by
which reader. We propose a new method, called wheel
graph, to solve the problem of the positioning region selection.
Through (6), we have
d estimate  10
  RSSI
10 n
.
(8)
Then equation (7) becomes equation (9), dest = destimate
2
 
 
d
  10n log actual  
d estimate   .(9)
10n
exp  
pd est  


2 2
ln 10d estimate 2




Fig. 8 depicts the PDF of the estimated distance.
From Fig. 8 it can be concluded that:
1) If keeping variance σ2 and actual distance dactual unchanged, the curve becomes concentrated and steep with
the path loss exponent n increasing.
2) If keeping variance σ2 and path loss factor n unchanged, the dispersion and smoothness with the actual
distance dactual increase.
3) If keeping actual distance dactual and path loss factor n unchanged, the curve shows that dispersion and
smoothness with the variance σ2 is increasing.
Fig. 9. Deployment of readers and tag.
Fig. 10. Wheel network of environment.
From the previous section, we know that every reader
will transmit RSSI information to the server. The server can
estimate the distance from tags to readers using (8). n1 is
the number of readers, the above situation can be described
by the following formula, this formula can be expressed as:
Reader _ n1 .
 Reader _1 Reader _ 2 Reader _ 3 Reader _ 4



d
d n1
d
d
d
3


1
2
4
After acquiring the distance information, we can establish
the wheel graph model in our environment. Fig. 10 shows
the wheel network of Fig. 9. As shown in Fig. 10, we connect all readers together. Each black or red dot represents
a reader node.
Then, range Dc = {d1 d2 d3 d4 … d5} from small value
to large one. And choose the smallest four values from it,
and meet the condition of d1，d2，d3，d4  Dc, given as
 Reader _1w Reader _ 2w Reader _ 3w Reader _ 4 w

,
d1w
d2w
d3w
d4w


Fig. 8. PDF of the estimated distance.
and d1w ≤ d2w ≤ d3w ≤ d4w. The following are the absolute
coordinates of the four readers,
Reader _1w  (x1w , y1w ) , Reader _ 2 w  (x 2 w , y2 w ) ,
Reader _ 3w  (x 3w , y3w ) , Reader _ 4 w  (x 4 w , y4 w ) .
2.4 Wheel Graph Model
In this section, we mainly discuss the positioning
environment and the choice of location area. The actual
positioning environment is too complex to use just three
readers for large-scale indoor localization. So we divide
a large area into many small square areas [14], as shown in
We can utilize the two shortest distances, d1w and d2w ,
to determine two closest readers near the object. Then we
regard the two readers as two vertices of the triangle. Next,
we use Reader_3w and Reader_4w respectively as the third
vertex to form a triangle, and judged by using the following
rules (Low of Cosines).
RADIOENGINEERING, VOL. 22, NO. 1, APRIL 2013
cos 1 
cos  2 
375
 x2 w  x1w  x3w  x1w    y2 w  y1w  y3w  y1w 
( x1w  x3 w )  ( y1w  y3 w )
2
2
( x1w  x2 w )  ( y1w  y2 w )
2
2
 x1w  x2 w  x4 w  x2 w    y1w  y2 w  y4 w  y2 w 
( x2 w  x4 w ) 2  ( y2 w  y4 w ) 2 ( x1w  x2 w ) 2  ( y1w  y2 w ) 2
When above formulas simultaneously satisfy the following conditions: cosθ1 > 0 & cosθ2 > 0, we take the point
as the third vertex. Using wheel graph method, we can determine the small area covered by three readers. And our
positioning region is a square area containing the triangle,
while the horizontal side of the triangle is the square area’s
one side. Fig. 11 displays the wheel graph of the whole
process. Fig. 12 shows the small square positioning area.
reader_3. There are four tags in the space, three reference
tags and one object tag. Suppose that during a period of
time, time slot 1 to time slot w, received signal strength
recorded as {RSSItj(i)} , j = 1,2,3, i = 1,2,3,4, t = 1,2,3…w.
Here, the {RSSItj(i)} denotes reader_j report the RSSIs
coming from tag_i during time slot 1 to time slot w. We can
calculate the average value of RSSIs, RSSI j (i ) denoted
as:
RSSI j (i )   j  10n j log(dist (i, j )) .
(10)
From (6) and (10), we can get the following formula.
RSSI j (i )  Pr (dist (i, j )) .
(11)
First, we calibrate the path loss exponent nj. We can
derive equation (12), (13) from (10) and (11). dist(i,j)
means the distance from tag_i to reader_j. We have:
RSSI j ( j )  a j  10n j log (dist ( j , j )) ,
(12)
RSSI j (i )  a j  10n j log (dist (i, j )) .
(13)
Then the difference of two equations, (12) and (13) is
Fig. 11. Wheel graph of the whole process.
RSSI j (i )  RSSI j ( j )  10n j log
dist (i, j ) .
dist ( j , j )
(14)
Then we can get the path loss exponent nj(i), which
denotes one calibration from reader_j to tag_i, and i  j. In
this section, j = 1,2,3, and i = 1,2,3.
n j (i) 
Fig. 12. Positioning area.
In the previous section, we have determined the positioning area. In order to calculate the probability in (7) or
(9), and the estimated distance in (8), we need to know the
propagation parameters α, n and σ. In our Bayesian localization algorithm, the on site reference tags and object tags
are used to calibrate the propagation parameters during the
process of localization. Every reader has a reference tag
deployed in the environment within its range. The reference
tags are static with known locations. After detection, the
readers report the values of RSSIs of tags to the server; the
server will process the received data and calibrate the propagation parameters. In the following part, we will explain
this process in detail.
The actual situation is shown in Fig. 12. There are
three readers in the specific area, reader_1, reader_2 and
(15)
For reader_j, the path loss factor is calculated twice,
we use its average value in our system.
nj 
2.5 Bayesian Localization Scheme
2.5.1 Calibration of the Propagation Parameters
RSSI j (i)  RSSI j (j) .
dist (i, j )
10 log
dist ( j , j )

n j (i )
i 1,2,3
2
.
(16)
nj denotes reader_j’s path loss factor to the environment.
Because there are three readers in our space, the path loss
factor should be calculated three times. Secondly, we calibrate the αj, from (10), we have
 j (i)  RSSI j (i )  10n j log (dist (i, j )) ,
j 
a j i 
,
i 1, 2 , 3 2

i  j  .
(17)
(18)
Finally, we calibrate the value of σ, as in formula (19).
Nr and Nt are the numbers of readers and tags. In our preset
situation, Nr = 3, Nt = 3, and std(RSSIj(i)) denotes the
standard deviation for vector [RSSI1j(i), RSSI2j(i),…,
RSSIwj(i)].
 
1
N r Nt
Nr Nt
[std (RSSI  i )] .
j
j 1 i 1
(19)
376
LONG XIAO, YE YIN, XIANGNONG WU, JIANWEI WANG, A LARGE-SCALE RF-BASED INDOOR LOCALIZATION SYSTEM …
2.5.2 Probability Localization Algorithm
As we know, the key of tracking and positioning is to
get location information. In our system, we define the location information x as a series of coordinate points. According to the Bayesian inference [15], we can obtain the following formula.
p  RSSI t |x  p  x 
p ( RSSI t | x) p ( x) (20)
t

p  x|RSSI  
t
p( RSSI t )
 p  RSSI |x  p  x  dx
sponding to the center of the grid’s coordinate, which is
defined as C(gxi,gyi), i  (1,2,3…N). In this area, small dots
represent tags, green ones are inference tags and red one is
object tag. Three big red circles are readers. A 2-dimensional or 3-dimensional map can be shown in Fig. 12 and
Fig. 13 respectively.
where p(x) is the priori probability of the object’s location.
p(RSSIt|x) can be derived from (7), which means the PDF
of RSSI under the situation of location information known.
p(RSSIt) is the marginal PDF of p(RSSIt,x). p(x|RSSIt) is the
key equation, we can obtain the location from it. Through
equation (20), we have following formula [16]
p ( x | RSSI t )  p ( RSSI t | x) p ( x) .
We assume that Bel(x) is the probability of object at
location x. Bel-(x) is the initial probability of object at
location x. β is a constant, to normalize the Bel(x) we have


Bel x    p RSSI xk Bel x  .
t

(21)
As described above, the initial p(x) could be set as 1/N,
thus the initial value of Bel(xk) is 1/N. The p(RSSIt|xk) is a
Fit Model, which can be calculated by the following
formula. Fig. 14 shows the Fit Model of a tracked point.
(22)
p( RSSI t | xk )   p( RSSI tj | xk ) .
Bel－(x) can be computed from equation (22).


Bel   x    p xk xk 1 p xk 1 RSSI t dxk 1
where xk is the location of object at time k and xk-1 is the
location of front moment. p(xk|xk-1) is the Dynamic Model
of system.
1
p( xk | xk 1 )  
0
xk  xk 1
xk  xk 1
(23)
For a static system, equation (22) can be converted
into equation (24).
Bel   xk   Bel  xk 1  .
(24)


3
(26)
j 1
There are two pictures in Fig. 14, the precondition of
this model is that RSSIs are known and unchanged. The
upper one represents the probability of the distance from
tag to one reader. The lower one represents the distances
from tag to the other two readers. From Fig. 14, we can
know that with the Fit Model relatively large values are
achieved, when the estimated locations are close to the
actual location. The maximum value is obtained when the
estimated location is actual location.
At the end of the algorithm, we evaluate the credibility of output by defining a confidence function Bel’(xk).
Then we can get our Bayesian estimation model.
Bel xk    p RSSI t xk Bel xk 1  .
Fig. 14. Fit Model of one tracked tag.
(25)
Bel  xk  
Bel xk 
.
maxBel  xk   xk , yk 
(27)
We set a threshold T(0 ≤ T ≤ 1) and only choose those
grids with Bel’(xk) > T. After one recursion, an estimated
area is obtained. We can achieve one more precise estimated area, if increasing the recursion times R.
3. System Performance Analysis
Fig. 13. Deployment environment.
As shown in Fig. 12, we divide the location area into
N grids. In our simulation, the grid size is 40 cm × 40 cm,
just like the size of a tile. Each grid has a coordinate corre-
This section is divided into three parts. First we discuss the average positioning distance error. Then we will
discuss the performance of the system with changing parameters. At last, we evaluate the system performance. As
we know, the performance will be affected by several undetermined parameters, for example, recursion time R, window size w and threshold T.
RADIOENGINEERING, VOL. 22, NO. 1, APRIL 2013
377
3.1 Average Localization Distance Error
In this part, we define ALE (Average Localization
Distance Error) to evaluate our system’s positioning error,
which is given by
ALE 
 DC  Bel x
 Bel x  C
gxk , gyk
k
k
 C gxk , gyk  
gxk , gyk 

.
area. For object 2 and object 3, the localization accuracy
increases much faster than that of object 1, with increasing
of the recursion time. Generally speaking, the position accuracy is increased with the increasing of the recursion
time.
(28)
The estimated points are in the estimated area, denoted as (gxk,gyk), k = 1,2,3,…E . E is the number of grids
covered by the estimated area. (x0,y0) is the actual coordinate of object tag. The D(C(gxk,gyk)) is denoted as the Euclid
Distance from estimated point to the object’s actual point.
We have


D C gxk , gyk  =
 gxk  x0    gyk  y0 
2
2
.
(29)
Fig. 15. Impact of Gaussian filter.
ALE is the average error distance between the object
and grid points in the estimated area. It represents the
accuracy of positioning. When the ALE is the minimum
value of zero, the system has the highest accuracy. The
greater the ALE value, the worse the position accuracy.
3.2 Impact of System Parameters
In this part, we will investigate the system performance depending on a number of system and environment
parameters and then simulate the experiment. We use
MATLAB to simulate the impact of the following parameters on the localization accuracy, which is represented by
average localization distance error (ALE): filtered data and
unfiltered data, recursion time R, threshold T, path loss
exponent n, standard deviation σ, and window size w, etc.
3.2.1 Impact of Filtered Data, Unfiltered Data and
Recursion Time
We study the impact of filtered data and unfiltered
data on average positioning error through experiment 1.
Experiment 1: We assume two scenarios: 1) the input
RSSI values are unfiltered data; 2) the input RSSI values
are filtered data. The system is based on the following configuration: grid size = 40 cm, threshold T = 0.5, path loss
exponent n = 3, standard deviation σ = 1.45, α = -59.7581,
window size w = 200.
Fig. 16. Tracked three objects.
3.2.2 Impact of Window Size
Experiment 2: The following system is configured as:
grid size = 40 cm, threshold T = 0.5, path loss exponent
n = 3, standard deviation σ = 1.45, α = -59.7581. Fig. 17
plots ALE as a function of the window size w. In this part,
we plan to evaluate the impact of window size on the localization accuracy. We perform the simulation for object 1,
object 2 and object 3 as depicted in Fig. 17.
From Fig. 17, we observe that ALE decreases as the
window size w increases. In other words, the greater the
window size w, the higher the localization accuracy. We
also find that the ALE decrease rate depends on the window
size. The ALE decrease rate will be very high when the
window size is relatively large.
Fig. 15 depicts ALE as a function of the recursion
time for the two scenarios. In the first scenario, we observe
that the decrease of ALE’s rate is very slow as we increase
the recursion time. After several rounds of recursion, ALE
is still large. But in the second scenario, we observe that
ALE decreases after increasing the recursion time. This is
due to the use of the Gaussian Filter which significantly
improves the reliability of RSSI values.
Fig. 16 illustrates ALE as a function of the recursion
time for different objects in the above second scenarios.
From Fig. 16, we observe that ALE reduced speed is
a function of the object locations within the basic detection
Fig. 17. Impact of window size.
378
LONG XIAO, YE YIN, XIANGNONG WU, JIANWEI WANG, A LARGE-SCALE RF-BASED INDOOR LOCALIZATION SYSTEM …
3.2.3 Impact of Threshold, Path Loss Exponent and
Standard Deviation
Experiment 3: The system is configured as: grid
size = 40 cm, path loss exponent n = 3, standard deviation
σ = 1.45, α = -59.7581. In this experiment, we aim to
evaluate the impact of threshold on the localization accuracy. We perform the simulation for object 1, object 2 and
object 3. Fig. 18 plots ALE as a function of the threshold T.
α = -59.7581. In this part, we evaluate the impact of standard deviation on the localization accuracy. Fig. 20 plots
ALE as a function of standard deviation σ.
From Fig. 20, we can see that localization accuracy
decreases with standard deviation σ increasing. The localization accuracy with small standard deviation is better.
From Fig. 18, it implies zero error for T = 1, this is
true only if no tags are found, i.e., with high threshold, the
error is low but almost no tags are identified. So, we must
have a principle, tags should be identified at the current
threshold, if no tags are found at the current threshold, we
discard the value.
Fig. 20. Impact of standard deviation.
3.3 System Performance
In this section, we test the system performance using
the following experiment.
Fig. 18. Impact of threshold.
Experiment 4: The system is configured as: grid
size = 40 cm, threshold T = 0.5, standard deviation
σ = 1.45, α = -59.7581. In this part, we plan to evaluate the
impact of path loss exponent on the localization accuracy.
We perform the simulation for object 1, object 2 and
object 3. Fig. 19 plots ALE as a function of the path loss
exponent n.
From (15) and (16), we can know that the path loss
exponent is not constant for the environment but a function
of both reader and tag position.
From Fig. 19, we can see that localization accuracy
increases with the increase of path loss exponent n. In
practical applications, the path loss exponent shouldn’t be
set manually but be set according to the environment.
When the environment path loss exponent is bigger, the
system positioning accuracy is higher.
Experiment 6: This experiment is designed to
probabilistically evaluate the system localization
performance under the following settings: Using the same
reader network topology as shown in Fig. 12, we track 169
objects which are uniformly distributed within the basic
detection area. The other simulation parameters are: grid
size = 40 cm, path loss exponent n = 3, standard deviation
σ = 1.45, α = -59.7581, threshold T = 0.5, 0.7, 0.9. We test
the system performance for two scenarios: 1) recursion
time R = 10; 2) recursion time R = 20.
Fig. 21 shows the cumulative Average Error Distance
distribution obtained from scenarios 1 and 2. We notice that
in over 90% percent of the objects the localization ALE
falls from within 50 cm. As expected, the increasing recursion time R will lead to higher localization accuracy. Furthermore, from Fig. 18, we also can get the conclusion that
in most of the situation, localization accuracy with large T
is better.
4. Conclusion
Fig. 19. Impact of path loss exponent.
Experiment 5: The system configuration: grid
size = 40 cm, path loss exponent n = 3, threshold T = 0.5,
In this paper, we presented a new method, an
easy-setup and cost-effective indoor localization method
based on off-the-shelf active RFID technology combined
with GF, WGM and PLA. Our system relies on a hierarchical architecture to cover an indoor environment. The proposed approach can calibrate the propagation parameters
according to the environment, reduce the uncertainty of
localization, and at the same time obtain high positioning
accuracy under conditions of large-scale indoor positioning.
Using MATLAB simulations we have evaluated the performance of our proposal. In the destined system settings,
the simulation results show that in 90% percent of the loca-
RADIOENGINEERING, VOL. 22, NO. 1, APRIL 2013
379
Fig. 21. 169 objects’ cumulative probability distribution.
lization estimation, the system provides objects location
with the ALE less than 50 cm, and in 70% percent of the
localization estimation, the system provides objects location with the ALE less than 20 cm. The simulation also
shows that the system performance improves with the
higher values of recursion time, window size, path loss
exponent and with the lower propagation standard deviation. The simulation results can prove that the proposed
system is an accuracy and cost-effective candidate for
indoor localization.
[5] Radio Frequency Identification (RFID) home page. [Online].
Available at: http://www.aimglobal.org/technologies/rfid/.
Acknowledgements
[9] MAJA, S., MLADEN, R., DINKO, B. RF Localization in indoor
environment. Radioengineering, 2012, vol. 21, no. 2, p. 557–567.
Authors are grateful to the anonymous reviewers for
their helpful comments. They also wish to note that the
research work is supported by the National Natural Science
Foundation of China (NO.61101209) and Shanghai Normal
University key project (A7001-12-002006).
References
[1] Garmin Corporation, About GPS. [Online]. Available at:
http://www.garmin.com/aboutGPS/.
[2] HUI, L., HOUSHANG, D., PAT, B., JING, L. Survey of wireless
indoor positioning techniques and systems. IEEE Transactions on
Systems, Man, and Cybernetics- Part C, 2007, vol. 37, no. 6,
p. 1067-1080.
[3] HIGHTOWER, J., BORRIELLO, G. A survey and taxonomy of
location sensing systems for ubiquitous computing. CSE 01-08-03,
University of Washington, Department of Computer Science and
Engineering, Seattle (USA). Technical Report, 2001.
[4] GUOQIANG, M., BARIS, F., BRIAN, D. O. A. Wireless sensor
network localization techniques. Computer Networks, 2007,
vol. 51, no. 10, p. 2529–2553.
[6] YE, Y., JUN, Z., JIAN, Y. Design of World Expo tour sites guide
system based on RFID technology. In Proceedings of 2010
International Conference on Audio Language and Image
Processing (ICALIP 2010). Shanghai (China), 2010, p. 1026-1030.
[7] HIGHTOWER, J., WANT, R., BORRIELLO, G. SpotON: An
indoor 3d location sensing technology based on RF signal strength.
UW-CSE 00-02-02, University of Washington, Department of
Computer Science and Engineering, Seattle (USA). Thesis, 2000.
[8] NI, L., LIU, Y., LAU, Y. C., PATIL, A. LANDMARC: Indoor
location sensing using active RFID. In Proceedings of the 1st
IEEE International Conference on Pervasive Computing and
Communications (PERCOM’03). Dallas (USA), 2003, p. 407-415.
[10] ZEPERNICK, H. J., WYSOCKI, T. A. Multipath channel
parameters for the indoor radio at 2.4 GHz ISMband. In
Proceedings of Vehicular Technology Conference, 1999 IEEE 49th.
Houston (USA), 1999, vol. 1, p. 190-193.
[11] ITO, K. Gaussian filter for nonlinear filtering problems. In
Proceedings of the 39th IEEE Conference on Decision and Control.
2000, vol. 2, p. 1218-1223.
[12] MINGHUI, Z., HUIQING, Z. Research on model of indoor
distance measurement based on receiving signal strength. In Proc.
of Internat. Conf. on Computer Design and Applications (ICCDA
2010). Qinhuangdao (China), 2010, vol. 5, p. 54-58.
[13] RAPPAPORT, T. S. Wireless Communications Principles and
Practices. Prentice-Hall Inc, 2002.
[14] WOOYONG, L., KYEONG, H., TAEYOUNG, K., DOODEOP,
E., JONGOK, K. Large scale indoor localization system based on
wireless sensor networks for ubiquitous computing. Wireless
Personal Communications, 2012, vol. 63, no. 1, p. 241-260.
[15] MADIGAN, D., ELNAHRAWY, E., MARTIN, R. Bayesian
indoor positioning systems. In Proceedings of 24th Annual Joint
Conference of the IEEE Computer and Communications Societies
(INFOCOM 2005). Miami (USA), 2005, vol. 2, p. 1217-1227.
[16] CHRISTOPHER, M. B. Pattern Recognition and Machine
Learning. New York: Springer-Verlag Inc, 2006.
380
LONG XIAO, YE YIN, XIANGNONG WU, JIANWEI WANG, A LARGE-SCALE RF-BASED INDOOR LOCALIZATION SYSTEM …
About Authors ...
Long XIAO was born in 1988 in Wuhan, China. He received his B.S. degree in Electronic and Information Engineering from the Huanggang Normal University in 2011.
He is now a graduate at Shanghai Normal University,
China. His research interests include wireless positioning,
wireless sensor network and signal processing. He has over
four pending patents.
Ye YIN was born in 1961 in Jiangsu province, China. He
received his M.S. and Ph.D degrees from the Tongji University. He is an associate professor of the College of Information, Mechanical and Electrical Engineering. His
research interests include relativity theory, information
theory, gravitational field communication, voice communication code.
Xiangnong WU received her B.S. and M.S. degrees from
the Huazhong University of Science and Technology,
China. She received Ph.D degree from the Nanyang Technological University, Singapore. She is an associate professor of the College of Information, Mechanical and Electrical Engineering. Her research interests include optical fiber
communication.
Jianwei WANG received his M.S. degrees from the Tongji
University. His research interests include RFID and signal
processing. He is the founder of Shanghai Zhen Zhuo Electrical Technology Ltd. Co.